Method and system for predicting influence of missense mutation on clinical efficacy based on deep learning
By combining drug molecule structure and mutant protein sequence features with deep learning models, the shortcomings of existing tools in terms of accuracy and generalization are solved, achieving high-precision and low-cost prediction of the effects of mutant drugs on efficacy and supporting high-throughput screening.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI JIAOTONG UNIV
- Filing Date
- 2023-04-11
- Publication Date
- 2026-06-02
AI Technical Summary
Existing tools for predicting the impact of mutations on the clinical efficacy of drugs are insufficient in terms of accuracy and generalization, and rely on expensive biological experiments and cannot explain model features, thus failing to effectively support the prediction of non-anti-cancer drugs and efficient mutation screening.
We employ deep learning methods, combining drug molecule structure and mutant protein sequence features, to construct a model using graph convolutional layers, fully connected layers, and Transformer blocks to predict the clinical efficacy impact of missense mutations. We use a self-attention mechanism to achieve interpretability and high-precision prediction.
Without relying on experimental information from crystal structure determination, a high-precision binary classification prediction of the impact of mutations on the clinical efficacy of drugs was achieved, supporting high-throughput mutation screening with low computational cost and interpretable model features.
Smart Images
Figure CN116403731B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of biology, and more specifically, to a method and system for predicting the impact of missense mutations on clinical drug efficacy based on deep learning. Background Technology
[0002] One of the greatest challenges of this century is precision medicine, particularly the exploration of personalized medicine for diseases. For example, in the field of cancer, targeted cancer therapy has been significantly correlated with improved survival rates over the past decade or so, thus becoming one of the standard strategies for cancer drug treatment. However, drug resistance occurs with targeted therapies for many diseases. Numerous studies have explored the mechanisms of resistance, suggesting that resistance mutations may be a significant contributing factor, altering the microenvironment and allowing the body to evade the effects of targeted drugs. It is noteworthy that finding successful treatment strategies requires diverse experimental conditions, cell lines, and various time-series modeling techniques, all of which are costly and time-consuming compared to traditional hypothesis-driven experimental methods. Therefore, developing computational methods based on big data to predict the impact of mutations on clinical efficacy can not only save computational costs but also provide guidance and early warning for clinical drug use.
[0003] In recent years, the public release of data from large-scale drug screening projects has promoted the development of precision medicine. Several tools have been released to predict the alteration and impact of mutations on the clinical efficacy of drugs. Existing tools can be broadly divided into two categories: one type predicts the change in protein-drug binding affinity (ΔΔG) caused by mutations by collecting data from biological experiments; this value is a continuous numerical value. The other type predicts whether mutations will change drug efficacy, i.e., whether a mutation is a drug-resistant mutation, by collecting labeled data from clinical literature; this approach models the problem as a binary classification. A typical example of the first type of tool is PremPLI, which uses a dataset of changes in protein-drug binding affinity (ΔΔG) caused by mutations measured in biological experiments to build a deep learning model for training and prediction. The second type of tool currently only has one example: Yu-Feng Lin et al., who collected data on anticancer drugs and, by acquiring and simulating the crystal structure after protein-drug docking, built an SVM machine learning framework for training and prediction.
[0004] In clinical applications, the lack of a unified threshold for judging changes in affinity makes it impossible to directly correlate predicted results with clinical efficacy. Therefore, binary classification is undoubtedly more direct and effective. However, acquiring and simulating the crystal structure after protein-drug docking undoubtedly increases computational costs. Deep learning has also proven to outperform machine learning in modeling complex data, and the prediction of mutational resistance in non-anticancer drugs is also worthy of study. Therefore, using deep learning methods to predict changes in drug clinical efficacy due to mutations, while minimizing restrictions on input information, is highly meaningful.
[0005] Predicting the impact of mutations on clinical efficacy using computational methods based on drug properties and protein sequences is extremely challenging. Different missense mutations are highly nuanced, requiring models to learn and distinguish the characteristics of different drug-mutation combinations. The problem of predicting the impact of mutations on clinical efficacy can be simply described as follows: given a drug and a missense mutation, based on the given drug's structural information (SMILE structure, PubChem substructure molecular fingerprint), and the sequences of the protein before and after the mutation, computational methods should predict whether the mutation will affect the drug's clinical efficacy. This problem is a binary classification problem.
[0006] With the rapid development of computing and storage capabilities, machine learning and deep learning are gradually demonstrating their ability to model and predict complex problems, such as in computer vision and natural language processing. Currently, the only tool capable of binary classification to predict whether mutations will affect clinical drug efficacy is the DRSP model, developed based on an anticancer drug dataset and published by Yu-Feng Lin et al. in 2022. This work makes predictions from the perspective of protein structure, using data from the Cancer Somatic Mutation Catalogue (COSMIC) database, whose data are collected from literature. The work uses a support vector machine (SVM) to model the data; if the crystal structure information of the drug-protein binding is unavailable, simulated docking structures are used. The test set used by this method includes three drug resistance mutations. The model trained on real crystal structures correctly predicted all three mutations, while the model trained on simulated crystal structures could only correctly predict one of them.
[0007] First, the existing binary classification prediction tool DRSP was specifically developed for anticancer drugs, and its predictive ability in other disease medication scenarios is limited. Second, due to limitations in model architecture and representation capabilities, the generalization and accuracy of the tool trained on new datasets cannot be guaranteed. Furthermore, DRSP relies too heavily on the crystal structure after protein-drug docking, while obtaining the accurate structure of the docking complex requires biological experiments; models simulating the structure of the docking complex lack sufficient accuracy, which contradicts our initial goal of developing a low-cost, high-precision tool. In addition, existing tools do not interpret the information learned by the model, making it impossible to know which features played a key role during training, which is detrimental to subsequent research and exploration of mutational drug resistance mechanisms.
[0008] Patent document CN107038351A discloses a method for systematically predicting the impact of omics variations on drug efficacy, including: 1) identifying three-dimensional functional sub-regions of proteins that determine drug efficacy through known clinically significant variations, and classifying gene / protein sites to establish a functional sub-region database T2; 2) comparing unknown clinically significant variations with the functional sub-region database T2, and predicting whether unknown clinically significant variations and known clinically significant variations have the same drug efficacy impact by analyzing whether the unknown clinically significant variations and known clinically significant variations belong to the same functional sub-region, the variation type, and the similarity of amino acid property changes. However, this invention does not use deep learning methods to perform high-precision prediction of the impact of mutations on clinical drug use while maintaining interpretability, and the input features are allowed to exclude clinical drug use information and crystal structure determination experimental information. Summary of the Invention
[0009] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for predicting the impact of missense mutations on clinical drug efficacy based on deep learning.
[0010] A method for predicting the impact of missense mutations on clinical efficacy based on deep learning, according to the present invention, includes:
[0011] Step S1: Obtain and preprocess the drug and mutant protein sequence dataset;
[0012] Step S2: Standardize the drug input data;
[0013] Step S3: Standardize the mutant protein sequence input data;
[0014] Step S4: Build a deep learning model based on the standardized data and output the predicted classification.
[0015] Preferably, in step S1:
[0016] Download all annotated entries on the effects of mutations on drug efficacy from the PharmGKB and Civic databases;
[0017] Step S1.1: Remove entries that are not single-point misspellings, remove entries with missing data, remove entries containing preset characters, remove entries with contradictory labels within the dataset and entries whose judgment difficulty is greater than the preset standard, and retain entries that have an impact and those that have no impact.
[0018] Step S1.2: Map mutations to the protein level by combining gene mutation and transcript information;
[0019] Step S1.3: Merge datasets from different sources. If duplicate entries are found, keep one. If entries with conflicting labels are found, remove those entries from the source datasets.
[0020] Step S1.4: The test set used by DRSP was removed from the dataset, and the labels were encoded as numbers for training and testing datasets: mutations that affect efficacy are assigned a value of 1, and mutations that do not affect efficacy are assigned a value of 0.
[0021] Preferably, in step S2:
[0022] The input drug name is matched using the PubChem API to obtain the drug's molecular structure information and substructure molecular fingerprints; the drug's molecular structure is then topologically encoded and generated into PyTorch format data.
[0023] The molecular fingerprint of a drug substructure is an 881-dimensional 0-1 vector that indicates whether the drug contains a specific substructure.
[0024] Preferably, in step S3:
[0025] The mutant protein sequence was downloaded from the UniProt protein database. The HHBlits features of the sequences before and after mutation were obtained in 30 dimensions. The tertiary structure of the protein was predicted using AlphaFold2 for the sequence before mutation. The secondary structure and relative solvent-accessible surface area of each residue were calculated. The secondary structure feature of each residue is a 3-dimensional 0-1 vector, and the relative solvent-accessible surface area feature of each residue is a 1-dimensional continuous value.
[0026] One-hot encoding of each residue in 20 dimensions was performed on the protein sequences before and after the mutation;
[0027] For all features of the protein before and after the mutation, feature slices were created with the mutation site as the center, the length of each pre-set amino acid before and after the mutation being retained, and the total length being a pre-set number of amino acids.
[0028] Preferably, in step S4:
[0029] The deep learning framework includes graph convolutional layers, fully connected layers, and Transformer blocks. Each Transformer block is implemented using a self-attention mechanism. The molecular structure of each input drug is graph-encoded and input into the graph convolutional layer for learning, forming the GCN branch. The molecular fingerprint of the drug is learned by a Transformer block, forming the molecular fingerprint branch. For mutation feature learning, the secondary structure features, relative solvent-accessible surface area features, and HHBlits features before and after each mutation are input into the same fully connected layer for learning, forming the sequence attribute branch. The sequence one-hot features before and after mutation are each learned by a separate Transformer block, forming the sequence one-hot branch.
[0030] In the second layer of the deep learning network, the output of the drug's molecular fingerprint branch and the sequence attribute branch are merged and input into the second fully connected layer for learning; the outputs of the two sequence One-Hot branches are merged and input into the second fully connected layer for learning.
[0031] In the third layer of the deep learning network, the outputs of two second fully connected layers and the GCN branch are merged and fed into the third fully connected layer for learning. The merged input is then processed by the SoftMax activation function for prediction and classification output.
[0032] A deep learning-based system for predicting the impact of missense mutations on clinical drug efficacy, according to the present invention, includes:
[0033] Module M1: Acquires and preprocesses drug and mutant protein sequence datasets;
[0034] Module M2: Standardizes drug input data;
[0035] Module M3: Normalizes the input data of mutant protein sequences;
[0036] Module M4: Builds a deep learning model based on the standardized data and outputs a predicted classification.
[0037] Preferably, in module M1:
[0038] Download all annotated entries on the effects of mutations on drug efficacy from the PharmGKB and Civic databases;
[0039] Module M1.1: Removes entries that are not single-point misspelled mutations, removes entries with missing data, removes entries containing preset characters, removes entries with contradictory labels within the dataset and entries whose judgment difficulty is greater than the preset standard, and retains entries that have an impact and those that have no impact.
[0040] Module M1.2: Combines gene mutation and transcript information to map mutations to the protein level;
[0041] Module M1.3: Merges datasets from different sources. If duplicate entries are found, one entry is retained. If entries with conflicting labels are found, the entries are removed from both sources.
[0042] Module M1.4: Removes the test set used by DRSP from the dataset and encodes the labels as numbers for training and testing datasets: mutations affect efficacy by 1, and mutations do not affect efficacy by 0.
[0043] Preferably, in module M2:
[0044] The input drug name is matched using the PubChem API to obtain the drug's molecular structure information and substructure molecular fingerprints; the drug's molecular structure is then topologically encoded and generated into PyTorch format data.
[0045] The molecular fingerprint of a drug substructure is an 881-dimensional 0-1 vector that indicates whether the drug contains a specific substructure.
[0046] Preferably, in module M3:
[0047] The mutant protein sequence was downloaded from the UniProt protein database. The HHBlits features of the sequences before and after mutation were obtained in 30 dimensions. The tertiary structure of the protein was predicted using AlphaFold2 for the sequence before mutation. The secondary structure and relative solvent-accessible surface area of each residue were calculated. The secondary structure feature of each residue is a 3-dimensional 0-1 vector, and the relative solvent-accessible surface area feature of each residue is a 1-dimensional continuous value.
[0048] One-hot encoding of each residue in 20 dimensions was performed on the protein sequences before and after the mutation;
[0049] For all features of the protein before and after the mutation, feature slices were created with the mutation site as the center, the length of each pre-set amino acid before and after the mutation being retained, and the total length being a pre-set number of amino acids.
[0050] Preferably, in module M4:
[0051] The deep learning framework includes graph convolutional layers, fully connected layers, and Transformer blocks. Each Transformer block is implemented using a self-attention mechanism. The molecular structure of each input drug is graph-encoded and input into the graph convolutional layer for learning, forming the GCN branch. The molecular fingerprint of the drug is learned by a Transformer block, forming the molecular fingerprint branch. For mutation feature learning, the secondary structure features, relative solvent-accessible surface area features, and HHBlits features before and after each mutation are input into the same fully connected layer for learning, forming the sequence attribute branch. The sequence one-hot features before and after mutation are each learned by a separate Transformer block, forming the sequence one-hot branch.
[0052] In the second layer of the deep learning network, the output of the drug's molecular fingerprint branch and the sequence attribute branch are merged and input into the second fully connected layer for learning; the outputs of the two sequence One-Hot branches are merged and input into the second fully connected layer for learning.
[0053] In the third layer of the deep learning network, the outputs of two second fully connected layers and the GCN branch are merged and fed into the third fully connected layer for learning. The merged input is then processed by the SoftMax activation function for prediction and classification output.
[0054] Compared with the prior art, the present invention has the following beneficial effects:
[0055] 1. This invention can predict the impact of mutations on clinical efficacy using only the structural properties of the drug itself and the sequence of the mutated protein;
[0056] 2. Because this invention uses fewer feature dimensions, it supports high-throughput prediction of the impact of mutations on drug efficacy with low computational overhead.
[0057] 3. This invention can use deep learning methods to predict the impact of mutations on clinical medication with high accuracy while maintaining interpretability, and the input features are allowed to exclude clinical medication information and crystal structure determination experimental information. Attached Figure Description
[0058] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:
[0059] Figure 1 This is a diagram illustrating the internal framework of the present invention, from input data to predicted output. Detailed Implementation
[0060] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.
[0061] Example 1:
[0062] According to the present invention, a method for predicting the impact of missense mutations on clinical drug efficacy based on deep learning is provided, such as... Figure 1 As shown, it includes:
[0063] Step S1: Obtain and preprocess the drug and mutant protein sequence dataset;
[0064] Specifically, in step S1:
[0065] Download all annotated entries on the effects of mutations on drug efficacy from the PharmGKB and Civic databases;
[0066] Step S1.1: Remove entries that are not single-point misspellings, remove entries with missing data, remove entries containing preset characters, remove entries with contradictory labels within the dataset and entries whose judgment difficulty is greater than the preset standard, and retain entries that have an impact and those that have no impact.
[0067] Step S1.2: Map mutations to the protein level by combining gene mutation and transcript information;
[0068] Step S1.3: Merge datasets from different sources. If duplicate entries are found, keep one. If entries with conflicting labels are found, remove those entries from the source datasets.
[0069] Step S1.4: The test set used by DRSP was removed from the dataset, and the labels were encoded as numbers for training and testing datasets: mutations that affect efficacy are assigned a value of 1, and mutations that do not affect efficacy are assigned a value of 0.
[0070] Step S2: Standardize the drug input data;
[0071] Specifically, in step S2:
[0072] The input drug name is matched using the PubChem API to obtain the drug's molecular structure information and substructure molecular fingerprints; the drug's molecular structure is then topologically encoded and generated into PyTorch format data.
[0073] The molecular fingerprint of a drug substructure is an 881-dimensional 0-1 vector that indicates whether the drug contains a specific substructure.
[0074] Step S3: Standardize the mutant protein sequence input data;
[0075] Specifically, in step S3:
[0076] The mutant protein sequence was downloaded from the UniProt protein database. The HHBlits features of the sequences before and after mutation were obtained in 30 dimensions. The tertiary structure of the protein was predicted using AlphaFold2 for the sequence before mutation. The secondary structure and relative solvent-accessible surface area of each residue were calculated. The secondary structure feature of each residue is a 3-dimensional 0-1 vector, and the relative solvent-accessible surface area feature of each residue is a 1-dimensional continuous value.
[0077] One-hot encoding of each residue in 20 dimensions was performed on the protein sequences before and after the mutation;
[0078] For all features of the protein before and after the mutation, feature slices were created with the mutation site as the center, the length of each pre-set amino acid before and after the mutation being retained, and the total length being a pre-set number of amino acids.
[0079] Step S4: Build a deep learning model based on the standardized data and output the predicted classification.
[0080] Specifically, in step S4:
[0081] The deep learning framework includes graph convolutional layers, fully connected layers, and Transformer blocks. Each Transformer block is implemented using a self-attention mechanism. The molecular structure of each input drug is graph-encoded and input into the graph convolutional layer for learning, forming the GCN branch. The molecular fingerprint of the drug is learned by a Transformer block, forming the molecular fingerprint branch. For mutation feature learning, the secondary structure features, relative solvent-accessible surface area features, and HHBlits features before and after each mutation are input into the same fully connected layer for learning, forming the sequence attribute branch. The sequence one-hot features before and after mutation are each learned by a separate Transformer block, forming the sequence one-hot branch.
[0082] In the second layer of the deep learning network, the output of the drug's molecular fingerprint branch and the sequence attribute branch are merged and input into the second fully connected layer for learning; the outputs of the two sequence One-Hot branches are merged and input into the second fully connected layer for learning.
[0083] In the third layer of the deep learning network, the outputs of two second fully connected layers and the GCN branch are merged and fed into the third fully connected layer for learning. The merged input is then processed by the SoftMax activation function for prediction and classification output.
[0084] Example 2:
[0085] Example 2 is a preferred embodiment of Example 1, and is used to illustrate the present invention in more detail.
[0086] The present invention also provides a deep learning-based system for predicting the impact of missense mutations on clinical efficacy. The deep learning-based system for predicting the impact of missense mutations on clinical efficacy can be implemented by executing the process steps of the deep learning-based method for predicting the impact of missense mutations on clinical efficacy. That is, those skilled in the art can understand the deep learning-based method for predicting the impact of missense mutations on clinical efficacy as a preferred embodiment of the deep learning-based system for predicting the impact of missense mutations on clinical efficacy.
[0087] A deep learning-based system for predicting the impact of missense mutations on clinical drug efficacy, according to the present invention, includes:
[0088] Module M1: Acquires and preprocesses drug and mutant protein sequence datasets;
[0089] Specifically, in module M1:
[0090] Download all annotated entries on the effects of mutations on drug efficacy from the PharmGKB and Civic databases;
[0091] Module M1.1: Removes entries that are not single-point misspelled mutations, removes entries with missing data, removes entries containing preset characters, removes entries with contradictory labels within the dataset and entries whose judgment difficulty is greater than the preset standard, and retains entries that have an impact and those that have no impact.
[0092] Module M1.2: Combines gene mutation and transcript information to map mutations to the protein level;
[0093] Module M1.3: Merges datasets from different sources. If duplicate entries are found, one entry is retained. If entries with conflicting labels are found, the entries are removed from both sources.
[0094] Module M1.4: Removes the test set used by DRSP from the dataset and encodes the labels as numbers for training and testing datasets: mutations affect efficacy by 1, and mutations do not affect efficacy by 0.
[0095] Module M2: Standardizes drug input data;
[0096] Specifically, in module M2:
[0097] The input drug name is matched using the PubChem API to obtain the drug's molecular structure information and substructure molecular fingerprints; the drug's molecular structure is then topologically encoded and generated into PyTorch format data.
[0098] The molecular fingerprint of a drug substructure is an 881-dimensional 0-1 vector that indicates whether the drug contains a specific substructure.
[0099] Module M3: Normalizes the input data of mutant protein sequences;
[0100] Specifically, in module M3:
[0101] The mutant protein sequence was downloaded from the UniProt protein database. The HHBlits features of the sequences before and after mutation were obtained in 30 dimensions. The tertiary structure of the protein was predicted using AlphaFold2 for the sequence before mutation. The secondary structure and relative solvent-accessible surface area of each residue were calculated. The secondary structure feature of each residue is a 3-dimensional 0-1 vector, and the relative solvent-accessible surface area feature of each residue is a 1-dimensional continuous value.
[0102] One-hot encoding of each residue in 20 dimensions was performed on the protein sequences before and after the mutation;
[0103] For all features of the protein before and after the mutation, feature slices were created with the mutation site as the center, the length of each pre-set amino acid before and after the mutation being retained, and the total length being a pre-set number of amino acids.
[0104] Module M4: Builds a deep learning model based on the standardized data and outputs a predicted classification.
[0105] Specifically, in module M4:
[0106] The deep learning framework includes graph convolutional layers, fully connected layers, and Transformer blocks. Each Transformer block is implemented using a self-attention mechanism. The molecular structure of each input drug is graph-encoded and input into the graph convolutional layer for learning, forming the GCN branch. The molecular fingerprint of the drug is learned by a Transformer block, forming the molecular fingerprint branch. For mutation feature learning, the secondary structure features, relative solvent-accessible surface area features, and HHBlits features before and after each mutation are input into the same fully connected layer for learning, forming the sequence attribute branch. The sequence one-hot features before and after mutation are each learned by a separate Transformer block, forming the sequence one-hot branch.
[0107] In the second layer of the deep learning network, the output of the drug's molecular fingerprint branch and the sequence attribute branch are merged and input into the second fully connected layer for learning; the outputs of the two sequence One-Hot branches are merged and input into the second fully connected layer for learning.
[0108] In the third layer of the deep learning network, the outputs of two second fully connected layers and the GCN branch are merged and fed into the third fully connected layer for learning. The merged input is then processed by the SoftMax activation function for prediction and classification output.
[0109] Example 3:
[0110] Example 3 is a preferred example of Example 1, and is used to illustrate the present invention in more detail.
[0111] This invention relates to the application of computer technology and computational biology in the field of drug resistance prediction, and uses deep learning methods to predict the impact of mutations on clinical efficacy based on drug property characteristics and protein sequence characteristics.
[0112] The purpose of this invention is to fill the gap in the general binary classification prediction of the impact of mutations on drug efficacy. Using only the structural information of the drug itself and the protein mutation sequence information, and allowing for feature gaps, the invention performs deep learning training on a clinical labeled dataset to make accurate and reliable predictions, and uses an attention mechanism to give the model interpretability.
[0113] Step 1: Obtain and preprocess the dataset
[0114] Training a deep learning model requires a non-redundant standard database. On December 21, 2022, we downloaded all labeled entries related to the effect of mutations on drug efficacy from the PharmGKB and Civic databases. First, we removed entries containing non-single-point missense mutations, entries with missing data, entries containing anomalous characters, entries with contradictory labels within the dataset, and entries that were difficult to determine, retaining only entries with or without impact. Second, we combined gene mutation and transcript information to map mutations to the protein level. Third, we merged the two datasets from different sources; if duplicate entries were found, only one was retained; if entries with contradictory labels were found, they were removed from both sources. Fourth, to ensure a fair comparison with the DRSP tool, we removed the test set used by DRSP from the dataset for subsequent external validation to evaluate the effectiveness of our invention. At this point, we retained 710 mutation-drug efficacy entries, encoding the labels as numbers (mutation affects efficacy: 1, mutation does not affect efficacy: 0), which served as the training and testing datasets for our invention.
[0115] Step 2: Standardization of Drug Input Data
[0116] This invention requires obtaining the structural features of the input drug. We match the drug name using PubChemAPI to obtain the drug's molecular structure SMILE information and substructure molecular fingerprint. The drug's molecular structure SMILE is topologically encoded using the RDKit package (developed in Python) and the PyTorch Geometric package and converted into PyTorch format data. The drug's substructure molecular fingerprint is an 881-dimensional 0-1 vector indicating whether the drug contains a specific substructure, as shown in the attached figure. Figure 1 As shown.
[0117] Step 3: Normalization of mutant protein sequence input data
[0118] This invention requires obtaining the protein sequences before and after mutation. We downloaded the sequence of the mutated protein from the UniProt protein database and input both sequences into HHBlits software to obtain 30-dimensional HHBlits features. Furthermore, we used AlphaFold2 software to predict the tertiary structure of the protein before and after mutation, and used DSSP software to calculate the secondary structure and relative accessible surface area (RLA) of each residue. The secondary structure feature of each residue is a 3-dimensional 0-1 vector, and the RLA feature is a 1-dimensional continuous value. Subsequently, we performed 20-dimensional one-hot encoding on each residue of the protein sequences before and after mutation. Finally, to standardize the input size of the protein sequences, for all features of the protein before and after mutation, we created feature slices centered on the mutation site, retaining 30 amino acids before and after the mutation site, for a total length of 61 amino acids, as shown in the attached figure. Figure 1 As shown. Through the above processing, each mutation has 61×20-dimensional One-Hot features before and after the mutation, 61×30-dimensional HHBlits features before and after the mutation, 61×3-dimensional secondary structure features before the mutation, and 61×1-dimensional relative solvent accessible surface area features before the mutation.
[0119] Step 4: Build a deep learning model
[0120] The deep learning framework of this invention consists of three important components, as shown in the appendix. Figure 1As shown, the layers are a Graph Convolutional Layer (GCN), a Fully-connected Layer, and a Transformer block. Each Transformer block is implemented using a self-attention mechanism. The molecular structure (SMILE) of each input drug is graph-encoded and fed into the GCN layer for learning (referred to as the GCN branch), while the drug's molecular fingerprint is learned by a Transformer block (referred to as the drug's molecular fingerprint branch). For mutation feature learning, the secondary structure features, relative solvent-accessible surface area features, and HHBlits features before and after each mutation are fed into the same fully connected layer for learning (referred to as the sequence attribute branch), while the sequence one-hot features before and after mutation are each learned by a separate Transformer block (referred to as the sequence one-hot branch). Subsequently, in the second layer of the deep learning network, the output of the drug's molecular fingerprint branch and the sequence attribute branch are merged and fed into the second fully connected layer for learning; the outputs of the two sequence one-hot features are also merged and fed into the second fully connected layer for learning. In the third layer of the network, we merge the outputs of the two second fully connected layers and the GCN branch, and then feed them into the third fully connected layer for learning. At this point, all the different input branches have been merged. Finally, the merged input is passed through the SoftMax activation function to predict the classification output.
[0121] The entire model framework, including code implementation, training, and testing, was implemented using the PyTorch deep learning framework. The entire experiment was conducted on an NVIDIA 3090 GPU. The batch size for model training was 128, and the learning rate was 0.0005. The main hyperparameters of the model (including the number of network layers, batch size, and learning rate) were optimized to achieve the best results on the validation set. The Adam optimizer was used to train the neural network parameters.
[0122] This invention can predict the impact of mutations on clinical efficacy using only the structural properties of the drug itself and the sequence of the mutated protein. To comprehensively evaluate the technical effectiveness of this invention, we used 130 items from the dataset as an independent test set and the three test items used in the DRSP as an external test set. After training and five-fold cross-validation, our invention achieved a prediction accuracy (ACC) of 0.84, an F1 score of 0.88, and an area under the receiver operating curve (AUC) of 0.91 on the independent test set. All predictions were correct on the external test set.
[0123] Because this invention uses fewer feature dimensions, it supports high-throughput prediction of the impact of mutations on drug efficacy with low computational overhead. Notably, this invention is currently the only one on the market that can use deep learning methods to predict the impact of mutations on clinical drug use with high accuracy while maintaining interpretability. Furthermore, the input features can exclude clinical drug use information and crystal structure determination experimental information, which is unique to this invention.
[0124] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.
[0125] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.
Claims
1. A method for predicting the impact of missense mutations on clinical drug efficacy based on deep learning, characterized in that, include: Step S1: Obtain and preprocess the drug and mutant protein sequence dataset; Step S2: Standardize the drug input data; Step S3: Standardize the mutant protein sequence input data; Step S4: Build a deep learning model based on the standardized data and output the predicted classification; In step S3: Download the mutant protein sequence from the UniProt protein database and obtain 30-dimensional HHBlits features from the sequences before and after mutation; The tertiary structure of the protein was predicted using AlphaFold2 based on the unmutated sequence, and the secondary structure and relative solvent-accessible surface area of each residue were calculated. The secondary structure feature of each residue is a 3-dimensional 0-1 vector, and the relative solvent-accessible surface area feature of each residue is a 1-dimensional continuous value. One-hot encoding of each residue in 20 dimensions was performed on the protein sequences before and after the mutation; For all features of the protein before and after the mutation, feature slices were created with the mutation site as the center, the length of each pre-set amino acid before and after the mutation being retained, and the total length being a pre-set number of amino acids. In step S4: The deep learning framework includes graph convolutional layers, fully connected layers, and Transformer blocks. Each Transformer block is implemented using a self-attention mechanism. The molecular structure of each input drug is graph-encoded and input into the graph convolutional layer for learning, forming the GCN branch. The molecular fingerprint of the drug is learned by a Transformer block, forming the molecular fingerprint branch. For mutation feature learning, the secondary structure features, relative solvent-accessible surface area features, and HHBlits features before and after each mutation are input into the same fully connected layer for learning, forming the sequence attribute branch. The sequence one-hot features before and after mutation are each learned by a separate Transformer block, forming the sequence one-hot branch. In the second layer of the deep learning network, the output of the drug's molecular fingerprint branch and the sequence attribute branch are merged and input into the second fully connected layer for learning; the outputs of the two sequence One-Hot branches are merged and input into the second fully connected layer for learning. In the third layer of the deep learning network, the outputs of two second fully connected layers and the GCN branch are merged and fed into the third fully connected layer for learning. The merged input is then processed by the SoftMax activation function for prediction and classification output.
2. The method for predicting the impact of missense mutations on clinical efficacy based on deep learning according to claim 1, characterized in that, In step S1: Download all annotated entries on the effects of mutations on drug efficacy from the PharmGKB and Civic databases; Step S1.1: Remove entries that are not single-point misspellings, remove entries with missing data, remove entries containing preset characters, remove entries with contradictory labels within the dataset and entries whose judgment difficulty is greater than the preset standard, and retain entries that have an impact and those that have no impact. Step S1.2: Map mutations to the protein level by combining gene mutation and transcript information; Step S1.3: Merge datasets from different sources. If duplicate entries are found, keep one. If entries with conflicting labels are found, remove those entries from the source datasets. Step S1.4: The test set used by DRSP was removed from the dataset, and the labels were encoded as numbers for training and testing datasets: mutations that affect efficacy are assigned a value of 1, and mutations that do not affect efficacy are assigned a value of 0.
3. The method for predicting the impact of missense mutations on clinical efficacy based on deep learning according to claim 1, characterized in that, In step S2: The input drug name is matched using the PubChem API to obtain the drug's molecular structure information and substructure molecular fingerprints; the drug's molecular structure is then topologically encoded and generated into PyTorch format data. The molecular fingerprint of a drug substructure is an 881-dimensional 0-1 vector that indicates whether the drug contains a specific substructure.
4. A deep learning-based system for predicting the impact of missense mutations on clinical drug efficacy, characterized in that, include: Module M1: Acquires and preprocesses drug and mutant protein sequence datasets; Module M2: Standardizes drug input data; Module M3: Normalizes the input data of mutant protein sequences; Module M4: Builds a deep learning model based on the standardized data and outputs a predicted classification. In module M3: Download the mutant protein sequence from the UniProt protein database and obtain 30-dimensional HHBlits features from the sequences before and after mutation; The tertiary structure of the protein was predicted using AlphaFold2 based on the unmutated sequence, and the secondary structure and relative solvent-accessible surface area of each residue were calculated. The secondary structure feature of each residue is a 3-dimensional 0-1 vector, and the relative solvent-accessible surface area feature of each residue is a 1-dimensional continuous value. One-hot encoding of each residue in 20 dimensions was performed on the protein sequences before and after the mutation; For all features of the protein before and after the mutation, feature slices were created with the mutation site as the center, the length of each pre-set amino acid before and after the mutation being retained, and the total length being a pre-set number of amino acids. In module M4: The deep learning framework includes graph convolutional layers, fully connected layers, and Transformer blocks. Each Transformer block is implemented using a self-attention mechanism. The molecular structure of each input drug is graph-encoded and input into the graph convolutional layer for learning, forming the GCN branch. The molecular fingerprint of the drug is learned by a Transformer block, forming the molecular fingerprint branch. For mutation feature learning, the secondary structure features, relative solvent-accessible surface area features, and HHBlits features before and after each mutation are input into the same fully connected layer for learning, forming the sequence attribute branch. The sequence one-hot features before and after mutation are each learned by a separate Transformer block, forming the sequence one-hot branch. In the second layer of the deep learning network, the output of the drug's molecular fingerprint branch and the sequence attribute branch are merged and input into the second fully connected layer for learning; the outputs of the two sequence One-Hot branches are merged and input into the second fully connected layer for learning. In the third layer of the deep learning network, the outputs of two second fully connected layers and the GCN branch are merged and fed into the third fully connected layer for learning. The merged input is then processed by the SoftMax activation function for prediction and classification output.
5. The deep learning-based missense mutation impact prediction system for clinical drug efficacy according to claim 4, characterized in that, In module M1: Download all annotated entries on the effects of mutations on drug efficacy from the PharmGKB and Civic databases; Module M1.1: Removes entries that are not single-point misspelled mutations, removes entries with missing data, removes entries containing preset characters, removes entries with contradictory labels within the dataset and entries whose judgment difficulty is greater than the preset standard, and retains entries that have an impact and those that have no impact. Module M1.2: Combines gene mutation and transcript information to map mutations to the protein level; Module M1.3: Merges datasets from different sources. If duplicate entries are found, one entry is retained. If entries with conflicting labels are found, the entries are removed from both sources. Module M1.4: Removes the test set used by DRSP from the dataset and encodes the labels as numbers for training and testing datasets: mutations affect efficacy by 1, and mutations do not affect efficacy by 0.
6. The deep learning-based missense mutation impact prediction system for clinical drug efficacy according to claim 4, characterized in that, In module M2: The input drug name is matched using the PubChem API to obtain the drug's molecular structure information and substructure molecular fingerprints; the drug's molecular structure is then topologically encoded and generated into PyTorch format data. The molecular fingerprint of a drug substructure is an 881-dimensional 0-1 vector that indicates whether the drug contains a specific substructure.